LeaseWeb nvidia-l4 DeepSeek-R1-8B FP16 decode tok/s FP16 (L4 GPU reference)
16.7 tok/s (decode, single-request)
Fields
id
benchmarks/lw-l4-deepseek8b-fp16-tgMetric
decode_tok_s
Value
16.7
Unit
tok/s (decode, single-request)
Model
DeepSeek-R1-8B
Methodology
LeaseWeb benchmark (HPE ProLiant DL385 Gen11, dual EPYC 9334 32C Zen4, 64GB DDR5; L4 reference single GPU). llama-bench + OpenLLM/llmperf. 512-token prompt, 128 generated, 24 CPU threads for LLM runs. decode tok/s FP16 (L4 GPU reference).
Measured by
third_party
Reproducible
yes
Thermal state
unknown
notes
Dual-socket EPYC 9334 (2x32C) unless L4. LLM at 24 threads.
Notes
Dual-socket EPYC 9334 (2x32C) unless L4. LLM at 24 threads.